Shanghai AI Lab's 8.9B "Next Concept Prediction" model hits same loss with 51.3% tokens

jiqizhixin · x · 2026-09-30

Shanghai AI Laboratory and SJTU's LUMIA Lab released NCP-ArchPreview, billed as the first 8.9B discrete latent-space foundation model, topping HuggingFace Daily Papers. It replaces pure Next-Token Prediction with "Next Concept Prediction" — predicting high-level concepts in a discrete latent space, echoing the pixel-to-latent shift in image generation and JEPA-style insights. Trained on Dolmo-3's 5.73T open-source corpus, it matches a full-budget model's final pretraining loss using only 51.3% of the token budget, converges 1.95x faster, and finishes 0.091 below the official baseline loss.

Original post →

More from Models

Models channel →